4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
52 * There is one ZIL per file system. Its on-disk (pool) format consists
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
69 * Disable intent logging replay. This global ZIL switch affects all pools.
71 int zil_replay_disable = 0;
72 SYSCTL_DECL(_vfs_zfs);
73 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
74 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
75 &zil_replay_disable, 0, "Disable intent logging replay");
78 * Tunable parameter for debugging or performance analysis. Setting
79 * zfs_nocacheflush will cause corruption on power loss if a volatile
80 * out-of-order write cache is enabled.
82 boolean_t zfs_nocacheflush = B_FALSE;
83 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
84 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
85 &zfs_nocacheflush, 0, "Disable cache flush");
86 boolean_t zfs_trim_enabled = B_TRUE;
87 SYSCTL_DECL(_vfs_zfs_trim);
88 TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled);
89 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
92 static kmem_cache_t *zil_lwb_cache;
94 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
96 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
97 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
101 * ziltest is by and large an ugly hack, but very useful in
102 * checking replay without tedious work.
103 * When running ziltest we want to keep all itx's and so maintain
104 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
105 * We subtract TXG_CONCURRENT_STATES to allow for common code.
107 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
110 zil_bp_compare(const void *x1, const void *x2)
112 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
113 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
115 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
117 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
120 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
122 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
129 zil_bp_tree_init(zilog_t *zilog)
131 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
132 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
136 zil_bp_tree_fini(zilog_t *zilog)
138 avl_tree_t *t = &zilog->zl_bp_tree;
142 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
143 kmem_free(zn, sizeof (zil_bp_node_t));
149 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
151 avl_tree_t *t = &zilog->zl_bp_tree;
156 if (BP_IS_EMBEDDED(bp))
159 dva = BP_IDENTITY(bp);
161 if (avl_find(t, dva, &where) != NULL)
162 return (SET_ERROR(EEXIST));
164 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
166 avl_insert(t, zn, where);
171 static zil_header_t *
172 zil_header_in_syncing_context(zilog_t *zilog)
174 return ((zil_header_t *)zilog->zl_header);
178 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
180 zio_cksum_t *zc = &bp->blk_cksum;
182 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
183 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
184 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
185 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
189 * Read a log block and make sure it's valid.
192 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
195 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
196 arc_flags_t aflags = ARC_FLAG_WAIT;
197 arc_buf_t *abuf = NULL;
201 if (zilog->zl_header->zh_claim_txg == 0)
202 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
204 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
205 zio_flags |= ZIO_FLAG_SPECULATIVE;
207 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
208 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
210 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
211 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
214 zio_cksum_t cksum = bp->blk_cksum;
217 * Validate the checksummed log block.
219 * Sequence numbers should be... sequential. The checksum
220 * verifier for the next block should be bp's checksum plus 1.
222 * Also check the log chain linkage and size used.
224 cksum.zc_word[ZIL_ZC_SEQ]++;
226 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
227 zil_chain_t *zilc = abuf->b_data;
228 char *lr = (char *)(zilc + 1);
229 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
231 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
232 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
233 error = SET_ERROR(ECKSUM);
235 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
237 *end = (char *)dst + len;
238 *nbp = zilc->zc_next_blk;
241 char *lr = abuf->b_data;
242 uint64_t size = BP_GET_LSIZE(bp);
243 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
245 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
246 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
247 (zilc->zc_nused > (size - sizeof (*zilc)))) {
248 error = SET_ERROR(ECKSUM);
250 ASSERT3U(zilc->zc_nused, <=,
251 SPA_OLD_MAXBLOCKSIZE);
252 bcopy(lr, dst, zilc->zc_nused);
253 *end = (char *)dst + zilc->zc_nused;
254 *nbp = zilc->zc_next_blk;
258 VERIFY(arc_buf_remove_ref(abuf, &abuf));
265 * Read a TX_WRITE log data block.
268 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
270 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
271 const blkptr_t *bp = &lr->lr_blkptr;
272 arc_flags_t aflags = ARC_FLAG_WAIT;
273 arc_buf_t *abuf = NULL;
277 if (BP_IS_HOLE(bp)) {
279 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
283 if (zilog->zl_header->zh_claim_txg == 0)
284 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
286 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
287 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
289 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
290 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
294 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
295 (void) arc_buf_remove_ref(abuf, &abuf);
302 * Parse the intent log, and call parse_func for each valid record within.
305 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
306 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
308 const zil_header_t *zh = zilog->zl_header;
309 boolean_t claimed = !!zh->zh_claim_txg;
310 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
311 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
312 uint64_t max_blk_seq = 0;
313 uint64_t max_lr_seq = 0;
314 uint64_t blk_count = 0;
315 uint64_t lr_count = 0;
316 blkptr_t blk, next_blk;
321 * Old logs didn't record the maximum zh_claim_lr_seq.
323 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
324 claim_lr_seq = UINT64_MAX;
327 * Starting at the block pointed to by zh_log we read the log chain.
328 * For each block in the chain we strongly check that block to
329 * ensure its validity. We stop when an invalid block is found.
330 * For each block pointer in the chain we call parse_blk_func().
331 * For each record in each valid block we call parse_lr_func().
332 * If the log has been claimed, stop if we encounter a sequence
333 * number greater than the highest claimed sequence number.
335 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
336 zil_bp_tree_init(zilog);
338 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
339 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
343 if (blk_seq > claim_blk_seq)
345 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
347 ASSERT3U(max_blk_seq, <, blk_seq);
348 max_blk_seq = blk_seq;
351 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
354 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
358 for (lrp = lrbuf; lrp < end; lrp += reclen) {
359 lr_t *lr = (lr_t *)lrp;
360 reclen = lr->lrc_reclen;
361 ASSERT3U(reclen, >=, sizeof (lr_t));
362 if (lr->lrc_seq > claim_lr_seq)
364 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
366 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
367 max_lr_seq = lr->lrc_seq;
372 zilog->zl_parse_error = error;
373 zilog->zl_parse_blk_seq = max_blk_seq;
374 zilog->zl_parse_lr_seq = max_lr_seq;
375 zilog->zl_parse_blk_count = blk_count;
376 zilog->zl_parse_lr_count = lr_count;
378 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
379 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
381 zil_bp_tree_fini(zilog);
382 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
388 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
391 * Claim log block if not already committed and not already claimed.
392 * If tx == NULL, just verify that the block is claimable.
394 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
395 zil_bp_tree_add(zilog, bp) != 0)
398 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
399 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
400 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
404 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
406 lr_write_t *lr = (lr_write_t *)lrc;
409 if (lrc->lrc_txtype != TX_WRITE)
413 * If the block is not readable, don't claim it. This can happen
414 * in normal operation when a log block is written to disk before
415 * some of the dmu_sync() blocks it points to. In this case, the
416 * transaction cannot have been committed to anyone (we would have
417 * waited for all writes to be stable first), so it is semantically
418 * correct to declare this the end of the log.
420 if (lr->lr_blkptr.blk_birth >= first_txg &&
421 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
423 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
428 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
430 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
436 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
438 lr_write_t *lr = (lr_write_t *)lrc;
439 blkptr_t *bp = &lr->lr_blkptr;
442 * If we previously claimed it, we need to free it.
444 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
445 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
447 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
453 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
457 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
458 lwb->lwb_zilog = zilog;
460 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
461 lwb->lwb_max_txg = txg;
464 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
465 lwb->lwb_nused = sizeof (zil_chain_t);
466 lwb->lwb_sz = BP_GET_LSIZE(bp);
469 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
472 mutex_enter(&zilog->zl_lock);
473 list_insert_tail(&zilog->zl_lwb_list, lwb);
474 mutex_exit(&zilog->zl_lock);
480 * Called when we create in-memory log transactions so that we know
481 * to cleanup the itxs at the end of spa_sync().
484 zilog_dirty(zilog_t *zilog, uint64_t txg)
486 dsl_pool_t *dp = zilog->zl_dmu_pool;
487 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
489 if (ds->ds_is_snapshot)
490 panic("dirtying snapshot!");
492 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
493 /* up the hold count until we can be written out */
494 dmu_buf_add_ref(ds->ds_dbuf, zilog);
499 zilog_is_dirty(zilog_t *zilog)
501 dsl_pool_t *dp = zilog->zl_dmu_pool;
503 for (int t = 0; t < TXG_SIZE; t++) {
504 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
511 * Create an on-disk intent log.
514 zil_create(zilog_t *zilog)
516 const zil_header_t *zh = zilog->zl_header;
524 * Wait for any previous destroy to complete.
526 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
528 ASSERT(zh->zh_claim_txg == 0);
529 ASSERT(zh->zh_replay_seq == 0);
534 * Allocate an initial log block if:
535 * - there isn't one already
536 * - the existing block is the wrong endianess
538 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
539 tx = dmu_tx_create(zilog->zl_os);
540 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
541 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
542 txg = dmu_tx_get_txg(tx);
544 if (!BP_IS_HOLE(&blk)) {
545 zio_free_zil(zilog->zl_spa, txg, &blk);
549 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
550 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
553 zil_init_log_chain(zilog, &blk);
557 * Allocate a log write buffer (lwb) for the first log block.
560 lwb = zil_alloc_lwb(zilog, &blk, txg);
563 * If we just allocated the first log block, commit our transaction
564 * and wait for zil_sync() to stuff the block poiner into zh_log.
565 * (zh is part of the MOS, so we cannot modify it in open context.)
569 txg_wait_synced(zilog->zl_dmu_pool, txg);
572 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
578 * In one tx, free all log blocks and clear the log header.
579 * If keep_first is set, then we're replaying a log with no content.
580 * We want to keep the first block, however, so that the first
581 * synchronous transaction doesn't require a txg_wait_synced()
582 * in zil_create(). We don't need to txg_wait_synced() here either
583 * when keep_first is set, because both zil_create() and zil_destroy()
584 * will wait for any in-progress destroys to complete.
587 zil_destroy(zilog_t *zilog, boolean_t keep_first)
589 const zil_header_t *zh = zilog->zl_header;
595 * Wait for any previous destroy to complete.
597 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
599 zilog->zl_old_header = *zh; /* debugging aid */
601 if (BP_IS_HOLE(&zh->zh_log))
604 tx = dmu_tx_create(zilog->zl_os);
605 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
606 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
607 txg = dmu_tx_get_txg(tx);
609 mutex_enter(&zilog->zl_lock);
611 ASSERT3U(zilog->zl_destroy_txg, <, txg);
612 zilog->zl_destroy_txg = txg;
613 zilog->zl_keep_first = keep_first;
615 if (!list_is_empty(&zilog->zl_lwb_list)) {
616 ASSERT(zh->zh_claim_txg == 0);
618 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
619 list_remove(&zilog->zl_lwb_list, lwb);
620 if (lwb->lwb_buf != NULL)
621 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
622 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
623 kmem_cache_free(zil_lwb_cache, lwb);
625 } else if (!keep_first) {
626 zil_destroy_sync(zilog, tx);
628 mutex_exit(&zilog->zl_lock);
634 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
636 ASSERT(list_is_empty(&zilog->zl_lwb_list));
637 (void) zil_parse(zilog, zil_free_log_block,
638 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
642 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
644 dmu_tx_t *tx = txarg;
645 uint64_t first_txg = dmu_tx_get_txg(tx);
651 error = dmu_objset_own_obj(dp, ds->ds_object,
652 DMU_OST_ANY, B_FALSE, FTAG, &os);
655 * EBUSY indicates that the objset is inconsistent, in which
656 * case it can not have a ZIL.
658 if (error != EBUSY) {
659 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
660 (unsigned long long)ds->ds_object, error);
665 zilog = dmu_objset_zil(os);
666 zh = zil_header_in_syncing_context(zilog);
668 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
669 if (!BP_IS_HOLE(&zh->zh_log))
670 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
671 BP_ZERO(&zh->zh_log);
672 dsl_dataset_dirty(dmu_objset_ds(os), tx);
673 dmu_objset_disown(os, FTAG);
678 * Claim all log blocks if we haven't already done so, and remember
679 * the highest claimed sequence number. This ensures that if we can
680 * read only part of the log now (e.g. due to a missing device),
681 * but we can read the entire log later, we will not try to replay
682 * or destroy beyond the last block we successfully claimed.
684 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
685 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
686 (void) zil_parse(zilog, zil_claim_log_block,
687 zil_claim_log_record, tx, first_txg);
688 zh->zh_claim_txg = first_txg;
689 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
690 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
691 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
692 zh->zh_flags |= ZIL_REPLAY_NEEDED;
693 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
694 dsl_dataset_dirty(dmu_objset_ds(os), tx);
697 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
698 dmu_objset_disown(os, FTAG);
703 * Check the log by walking the log chain.
704 * Checksum errors are ok as they indicate the end of the chain.
705 * Any other error (no device or read failure) returns an error.
709 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
718 error = dmu_objset_from_ds(ds, &os);
720 cmn_err(CE_WARN, "can't open objset %llu, error %d",
721 (unsigned long long)ds->ds_object, error);
725 zilog = dmu_objset_zil(os);
726 bp = (blkptr_t *)&zilog->zl_header->zh_log;
729 * Check the first block and determine if it's on a log device
730 * which may have been removed or faulted prior to loading this
731 * pool. If so, there's no point in checking the rest of the log
732 * as its content should have already been synced to the pool.
734 if (!BP_IS_HOLE(bp)) {
736 boolean_t valid = B_TRUE;
738 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
739 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
740 if (vd->vdev_islog && vdev_is_dead(vd))
741 valid = vdev_log_state_valid(vd);
742 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
749 * Because tx == NULL, zil_claim_log_block() will not actually claim
750 * any blocks, but just determine whether it is possible to do so.
751 * In addition to checking the log chain, zil_claim_log_block()
752 * will invoke zio_claim() with a done func of spa_claim_notify(),
753 * which will update spa_max_claim_txg. See spa_load() for details.
755 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
756 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
758 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
762 zil_vdev_compare(const void *x1, const void *x2)
764 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
765 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
776 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
778 avl_tree_t *t = &zilog->zl_vdev_tree;
780 zil_vdev_node_t *zv, zvsearch;
781 int ndvas = BP_GET_NDVAS(bp);
784 if (zfs_nocacheflush)
787 ASSERT(zilog->zl_writer);
790 * Even though we're zl_writer, we still need a lock because the
791 * zl_get_data() callbacks may have dmu_sync() done callbacks
792 * that will run concurrently.
794 mutex_enter(&zilog->zl_vdev_lock);
795 for (i = 0; i < ndvas; i++) {
796 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
797 if (avl_find(t, &zvsearch, &where) == NULL) {
798 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
799 zv->zv_vdev = zvsearch.zv_vdev;
800 avl_insert(t, zv, where);
803 mutex_exit(&zilog->zl_vdev_lock);
807 zil_flush_vdevs(zilog_t *zilog)
809 spa_t *spa = zilog->zl_spa;
810 avl_tree_t *t = &zilog->zl_vdev_tree;
815 ASSERT(zilog->zl_writer);
818 * We don't need zl_vdev_lock here because we're the zl_writer,
819 * and all zl_get_data() callbacks are done.
821 if (avl_numnodes(t) == 0)
824 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
826 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
828 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
829 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
832 kmem_free(zv, sizeof (*zv));
836 * Wait for all the flushes to complete. Not all devices actually
837 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
839 (void) zio_wait(zio);
841 spa_config_exit(spa, SCL_STATE, FTAG);
845 * Function called when a log block write completes
848 zil_lwb_write_done(zio_t *zio)
850 lwb_t *lwb = zio->io_private;
851 zilog_t *zilog = lwb->lwb_zilog;
852 dmu_tx_t *tx = lwb->lwb_tx;
854 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
855 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
856 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
857 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
858 ASSERT(!BP_IS_GANG(zio->io_bp));
859 ASSERT(!BP_IS_HOLE(zio->io_bp));
860 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
863 * Ensure the lwb buffer pointer is cleared before releasing
864 * the txg. If we have had an allocation failure and
865 * the txg is waiting to sync then we want want zil_sync()
866 * to remove the lwb so that it's not picked up as the next new
867 * one in zil_commit_writer(). zil_sync() will only remove
868 * the lwb if lwb_buf is null.
870 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
871 mutex_enter(&zilog->zl_lock);
874 mutex_exit(&zilog->zl_lock);
877 * Now that we've written this log block, we have a stable pointer
878 * to the next block in the chain, so it's OK to let the txg in
879 * which we allocated the next block sync.
885 * Initialize the io for a log block.
888 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
892 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
893 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
894 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
896 if (zilog->zl_root_zio == NULL) {
897 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
900 if (lwb->lwb_zio == NULL) {
901 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
902 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
903 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
904 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
909 * Define a limited set of intent log block sizes.
911 * These must be a multiple of 4KB. Note only the amount used (again
912 * aligned to 4KB) actually gets written. However, we can't always just
913 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
915 uint64_t zil_block_buckets[] = {
916 4096, /* non TX_WRITE */
917 8192+4096, /* data base */
918 32*1024 + 4096, /* NFS writes */
923 * Use the slog as long as the logbias is 'latency' and the current commit size
924 * is less than the limit or the total list size is less than 2X the limit.
925 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
927 uint64_t zil_slog_limit = 1024 * 1024;
928 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
929 (((zilog)->zl_cur_used < zil_slog_limit) || \
930 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
933 * Start a log block write and advance to the next log block.
934 * Calls are serialized.
937 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
941 spa_t *spa = zilog->zl_spa;
945 uint64_t zil_blksz, wsz;
948 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
949 zilc = (zil_chain_t *)lwb->lwb_buf;
950 bp = &zilc->zc_next_blk;
952 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
953 bp = &zilc->zc_next_blk;
956 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
959 * Allocate the next block and save its address in this block
960 * before writing it in order to establish the log chain.
961 * Note that if the allocation of nlwb synced before we wrote
962 * the block that points at it (lwb), we'd leak it if we crashed.
963 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
964 * We dirty the dataset to ensure that zil_sync() will be called
965 * to clean up in the event of allocation failure or I/O failure.
967 tx = dmu_tx_create(zilog->zl_os);
968 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
969 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
970 txg = dmu_tx_get_txg(tx);
975 * Log blocks are pre-allocated. Here we select the size of the next
976 * block, based on size used in the last block.
977 * - first find the smallest bucket that will fit the block from a
978 * limited set of block sizes. This is because it's faster to write
979 * blocks allocated from the same metaslab as they are adjacent or
981 * - next find the maximum from the new suggested size and an array of
982 * previous sizes. This lessens a picket fence effect of wrongly
983 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
986 * Note we only write what is used, but we can't just allocate
987 * the maximum block size because we can exhaust the available
990 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
991 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
993 zil_blksz = zil_block_buckets[i];
994 if (zil_blksz == UINT64_MAX)
995 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
996 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
997 for (i = 0; i < ZIL_PREV_BLKS; i++)
998 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
999 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1002 /* pass the old blkptr in order to spread log blocks across devs */
1003 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1006 ASSERT3U(bp->blk_birth, ==, txg);
1007 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1008 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1011 * Allocate a new log write buffer (lwb).
1013 nlwb = zil_alloc_lwb(zilog, bp, txg);
1015 /* Record the block for later vdev flushing */
1016 zil_add_block(zilog, &lwb->lwb_blk);
1019 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1020 /* For Slim ZIL only write what is used. */
1021 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1022 ASSERT3U(wsz, <=, lwb->lwb_sz);
1023 zio_shrink(lwb->lwb_zio, wsz);
1030 zilc->zc_nused = lwb->lwb_nused;
1031 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1034 * clear unused data for security
1036 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1038 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1041 * If there was an allocation failure then nlwb will be null which
1042 * forces a txg_wait_synced().
1048 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1050 lr_t *lrc = &itx->itx_lr; /* common log record */
1051 lr_write_t *lrw = (lr_write_t *)lrc;
1053 uint64_t txg = lrc->lrc_txg;
1054 uint64_t reclen = lrc->lrc_reclen;
1060 ASSERT(lwb->lwb_buf != NULL);
1061 ASSERT(zilog_is_dirty(zilog) ||
1062 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1064 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1065 dlen = P2ROUNDUP_TYPED(
1066 lrw->lr_length, sizeof (uint64_t), uint64_t);
1068 zilog->zl_cur_used += (reclen + dlen);
1070 zil_lwb_write_init(zilog, lwb);
1073 * If this record won't fit in the current log block, start a new one.
1075 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1076 lwb = zil_lwb_write_start(zilog, lwb);
1079 zil_lwb_write_init(zilog, lwb);
1080 ASSERT(LWB_EMPTY(lwb));
1081 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1082 txg_wait_synced(zilog->zl_dmu_pool, txg);
1087 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1088 bcopy(lrc, lr_buf, reclen);
1089 lrc = (lr_t *)lr_buf;
1090 lrw = (lr_write_t *)lrc;
1093 * If it's a write, fetch the data or get its blkptr as appropriate.
1095 if (lrc->lrc_txtype == TX_WRITE) {
1096 if (txg > spa_freeze_txg(zilog->zl_spa))
1097 txg_wait_synced(zilog->zl_dmu_pool, txg);
1098 if (itx->itx_wr_state != WR_COPIED) {
1103 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1104 dbuf = lr_buf + reclen;
1105 lrw->lr_common.lrc_reclen += dlen;
1107 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1110 error = zilog->zl_get_data(
1111 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1113 txg_wait_synced(zilog->zl_dmu_pool, txg);
1117 ASSERT(error == ENOENT || error == EEXIST ||
1125 * We're actually making an entry, so update lrc_seq to be the
1126 * log record sequence number. Note that this is generally not
1127 * equal to the itx sequence number because not all transactions
1128 * are synchronous, and sometimes spa_sync() gets there first.
1130 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1131 lwb->lwb_nused += reclen + dlen;
1132 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1133 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1134 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1140 zil_itx_create(uint64_t txtype, size_t lrsize)
1144 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1146 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1147 itx->itx_lr.lrc_txtype = txtype;
1148 itx->itx_lr.lrc_reclen = lrsize;
1149 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1150 itx->itx_lr.lrc_seq = 0; /* defensive */
1151 itx->itx_sync = B_TRUE; /* default is synchronous */
1157 zil_itx_destroy(itx_t *itx)
1159 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1163 * Free up the sync and async itxs. The itxs_t has already been detached
1164 * so no locks are needed.
1167 zil_itxg_clean(itxs_t *itxs)
1173 itx_async_node_t *ian;
1175 list = &itxs->i_sync_list;
1176 while ((itx = list_head(list)) != NULL) {
1177 list_remove(list, itx);
1178 kmem_free(itx, offsetof(itx_t, itx_lr) +
1179 itx->itx_lr.lrc_reclen);
1183 t = &itxs->i_async_tree;
1184 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1185 list = &ian->ia_list;
1186 while ((itx = list_head(list)) != NULL) {
1187 list_remove(list, itx);
1188 kmem_free(itx, offsetof(itx_t, itx_lr) +
1189 itx->itx_lr.lrc_reclen);
1192 kmem_free(ian, sizeof (itx_async_node_t));
1196 kmem_free(itxs, sizeof (itxs_t));
1200 zil_aitx_compare(const void *x1, const void *x2)
1202 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1203 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1214 * Remove all async itx with the given oid.
1217 zil_remove_async(zilog_t *zilog, uint64_t oid)
1220 itx_async_node_t *ian;
1227 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1229 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1232 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1234 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1235 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1237 mutex_enter(&itxg->itxg_lock);
1238 if (itxg->itxg_txg != txg) {
1239 mutex_exit(&itxg->itxg_lock);
1244 * Locate the object node and append its list.
1246 t = &itxg->itxg_itxs->i_async_tree;
1247 ian = avl_find(t, &oid, &where);
1249 list_move_tail(&clean_list, &ian->ia_list);
1250 mutex_exit(&itxg->itxg_lock);
1252 while ((itx = list_head(&clean_list)) != NULL) {
1253 list_remove(&clean_list, itx);
1254 kmem_free(itx, offsetof(itx_t, itx_lr) +
1255 itx->itx_lr.lrc_reclen);
1257 list_destroy(&clean_list);
1261 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1265 itxs_t *itxs, *clean = NULL;
1268 * Object ids can be re-instantiated in the next txg so
1269 * remove any async transactions to avoid future leaks.
1270 * This can happen if a fsync occurs on the re-instantiated
1271 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1272 * the new file data and flushes a write record for the old object.
1274 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1275 zil_remove_async(zilog, itx->itx_oid);
1278 * Ensure the data of a renamed file is committed before the rename.
1280 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1281 zil_async_to_sync(zilog, itx->itx_oid);
1283 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1286 txg = dmu_tx_get_txg(tx);
1288 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1289 mutex_enter(&itxg->itxg_lock);
1290 itxs = itxg->itxg_itxs;
1291 if (itxg->itxg_txg != txg) {
1294 * The zil_clean callback hasn't got around to cleaning
1295 * this itxg. Save the itxs for release below.
1296 * This should be rare.
1298 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1300 clean = itxg->itxg_itxs;
1302 ASSERT(itxg->itxg_sod == 0);
1303 itxg->itxg_txg = txg;
1304 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1306 list_create(&itxs->i_sync_list, sizeof (itx_t),
1307 offsetof(itx_t, itx_node));
1308 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1309 sizeof (itx_async_node_t),
1310 offsetof(itx_async_node_t, ia_node));
1312 if (itx->itx_sync) {
1313 list_insert_tail(&itxs->i_sync_list, itx);
1314 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1315 itxg->itxg_sod += itx->itx_sod;
1317 avl_tree_t *t = &itxs->i_async_tree;
1318 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1319 itx_async_node_t *ian;
1322 ian = avl_find(t, &foid, &where);
1324 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1325 list_create(&ian->ia_list, sizeof (itx_t),
1326 offsetof(itx_t, itx_node));
1327 ian->ia_foid = foid;
1328 avl_insert(t, ian, where);
1330 list_insert_tail(&ian->ia_list, itx);
1333 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1334 zilog_dirty(zilog, txg);
1335 mutex_exit(&itxg->itxg_lock);
1337 /* Release the old itxs now we've dropped the lock */
1339 zil_itxg_clean(clean);
1343 * If there are any in-memory intent log transactions which have now been
1344 * synced then start up a taskq to free them. We should only do this after we
1345 * have written out the uberblocks (i.e. txg has been comitted) so that
1346 * don't inadvertently clean out in-memory log records that would be required
1350 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1352 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1355 mutex_enter(&itxg->itxg_lock);
1356 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1357 mutex_exit(&itxg->itxg_lock);
1360 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1361 ASSERT(itxg->itxg_txg != 0);
1362 ASSERT(zilog->zl_clean_taskq != NULL);
1363 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1365 clean_me = itxg->itxg_itxs;
1366 itxg->itxg_itxs = NULL;
1368 mutex_exit(&itxg->itxg_lock);
1370 * Preferably start a task queue to free up the old itxs but
1371 * if taskq_dispatch can't allocate resources to do that then
1372 * free it in-line. This should be rare. Note, using TQ_SLEEP
1373 * created a bad performance problem.
1375 if (taskq_dispatch(zilog->zl_clean_taskq,
1376 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1377 zil_itxg_clean(clean_me);
1381 * Get the list of itxs to commit into zl_itx_commit_list.
1384 zil_get_commit_list(zilog_t *zilog)
1387 list_t *commit_list = &zilog->zl_itx_commit_list;
1388 uint64_t push_sod = 0;
1390 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1393 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1395 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1396 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1398 mutex_enter(&itxg->itxg_lock);
1399 if (itxg->itxg_txg != txg) {
1400 mutex_exit(&itxg->itxg_lock);
1404 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1405 push_sod += itxg->itxg_sod;
1408 mutex_exit(&itxg->itxg_lock);
1410 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1414 * Move the async itxs for a specified object to commit into sync lists.
1417 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1420 itx_async_node_t *ian;
1424 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1427 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1429 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1430 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1432 mutex_enter(&itxg->itxg_lock);
1433 if (itxg->itxg_txg != txg) {
1434 mutex_exit(&itxg->itxg_lock);
1439 * If a foid is specified then find that node and append its
1440 * list. Otherwise walk the tree appending all the lists
1441 * to the sync list. We add to the end rather than the
1442 * beginning to ensure the create has happened.
1444 t = &itxg->itxg_itxs->i_async_tree;
1446 ian = avl_find(t, &foid, &where);
1448 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1452 void *cookie = NULL;
1454 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1455 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1457 list_destroy(&ian->ia_list);
1458 kmem_free(ian, sizeof (itx_async_node_t));
1461 mutex_exit(&itxg->itxg_lock);
1466 zil_commit_writer(zilog_t *zilog)
1471 spa_t *spa = zilog->zl_spa;
1474 ASSERT(zilog->zl_root_zio == NULL);
1476 mutex_exit(&zilog->zl_lock);
1478 zil_get_commit_list(zilog);
1481 * Return if there's nothing to commit before we dirty the fs by
1482 * calling zil_create().
1484 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1485 mutex_enter(&zilog->zl_lock);
1489 if (zilog->zl_suspend) {
1492 lwb = list_tail(&zilog->zl_lwb_list);
1494 lwb = zil_create(zilog);
1497 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1498 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1499 txg = itx->itx_lr.lrc_txg;
1502 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1503 lwb = zil_lwb_commit(zilog, itx, lwb);
1504 list_remove(&zilog->zl_itx_commit_list, itx);
1505 kmem_free(itx, offsetof(itx_t, itx_lr)
1506 + itx->itx_lr.lrc_reclen);
1508 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1510 /* write the last block out */
1511 if (lwb != NULL && lwb->lwb_zio != NULL)
1512 lwb = zil_lwb_write_start(zilog, lwb);
1514 zilog->zl_cur_used = 0;
1517 * Wait if necessary for the log blocks to be on stable storage.
1519 if (zilog->zl_root_zio) {
1520 error = zio_wait(zilog->zl_root_zio);
1521 zilog->zl_root_zio = NULL;
1522 zil_flush_vdevs(zilog);
1525 if (error || lwb == NULL)
1526 txg_wait_synced(zilog->zl_dmu_pool, 0);
1528 mutex_enter(&zilog->zl_lock);
1531 * Remember the highest committed log sequence number for ztest.
1532 * We only update this value when all the log writes succeeded,
1533 * because ztest wants to ASSERT that it got the whole log chain.
1535 if (error == 0 && lwb != NULL)
1536 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1540 * Commit zfs transactions to stable storage.
1541 * If foid is 0 push out all transactions, otherwise push only those
1542 * for that object or might reference that object.
1544 * itxs are committed in batches. In a heavily stressed zil there will be
1545 * a commit writer thread who is writing out a bunch of itxs to the log
1546 * for a set of committing threads (cthreads) in the same batch as the writer.
1547 * Those cthreads are all waiting on the same cv for that batch.
1549 * There will also be a different and growing batch of threads that are
1550 * waiting to commit (qthreads). When the committing batch completes
1551 * a transition occurs such that the cthreads exit and the qthreads become
1552 * cthreads. One of the new cthreads becomes the writer thread for the
1553 * batch. Any new threads arriving become new qthreads.
1555 * Only 2 condition variables are needed and there's no transition
1556 * between the two cvs needed. They just flip-flop between qthreads
1559 * Using this scheme we can efficiently wakeup up only those threads
1560 * that have been committed.
1563 zil_commit(zilog_t *zilog, uint64_t foid)
1567 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1570 /* move the async itxs for the foid to the sync queues */
1571 zil_async_to_sync(zilog, foid);
1573 mutex_enter(&zilog->zl_lock);
1574 mybatch = zilog->zl_next_batch;
1575 while (zilog->zl_writer) {
1576 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1577 if (mybatch <= zilog->zl_com_batch) {
1578 mutex_exit(&zilog->zl_lock);
1583 zilog->zl_next_batch++;
1584 zilog->zl_writer = B_TRUE;
1585 zil_commit_writer(zilog);
1586 zilog->zl_com_batch = mybatch;
1587 zilog->zl_writer = B_FALSE;
1588 mutex_exit(&zilog->zl_lock);
1590 /* wake up one thread to become the next writer */
1591 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1593 /* wake up all threads waiting for this batch to be committed */
1594 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1598 * Called in syncing context to free committed log blocks and update log header.
1601 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1603 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1604 uint64_t txg = dmu_tx_get_txg(tx);
1605 spa_t *spa = zilog->zl_spa;
1606 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1610 * We don't zero out zl_destroy_txg, so make sure we don't try
1611 * to destroy it twice.
1613 if (spa_sync_pass(spa) != 1)
1616 mutex_enter(&zilog->zl_lock);
1618 ASSERT(zilog->zl_stop_sync == 0);
1620 if (*replayed_seq != 0) {
1621 ASSERT(zh->zh_replay_seq < *replayed_seq);
1622 zh->zh_replay_seq = *replayed_seq;
1626 if (zilog->zl_destroy_txg == txg) {
1627 blkptr_t blk = zh->zh_log;
1629 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1631 bzero(zh, sizeof (zil_header_t));
1632 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1634 if (zilog->zl_keep_first) {
1636 * If this block was part of log chain that couldn't
1637 * be claimed because a device was missing during
1638 * zil_claim(), but that device later returns,
1639 * then this block could erroneously appear valid.
1640 * To guard against this, assign a new GUID to the new
1641 * log chain so it doesn't matter what blk points to.
1643 zil_init_log_chain(zilog, &blk);
1648 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1649 zh->zh_log = lwb->lwb_blk;
1650 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1652 list_remove(&zilog->zl_lwb_list, lwb);
1653 zio_free_zil(spa, txg, &lwb->lwb_blk);
1654 kmem_cache_free(zil_lwb_cache, lwb);
1657 * If we don't have anything left in the lwb list then
1658 * we've had an allocation failure and we need to zero
1659 * out the zil_header blkptr so that we don't end
1660 * up freeing the same block twice.
1662 if (list_head(&zilog->zl_lwb_list) == NULL)
1663 BP_ZERO(&zh->zh_log);
1665 mutex_exit(&zilog->zl_lock);
1671 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1672 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1678 kmem_cache_destroy(zil_lwb_cache);
1682 zil_set_sync(zilog_t *zilog, uint64_t sync)
1684 zilog->zl_sync = sync;
1688 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1690 zilog->zl_logbias = logbias;
1694 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1698 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1700 zilog->zl_header = zh_phys;
1702 zilog->zl_spa = dmu_objset_spa(os);
1703 zilog->zl_dmu_pool = dmu_objset_pool(os);
1704 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1705 zilog->zl_logbias = dmu_objset_logbias(os);
1706 zilog->zl_sync = dmu_objset_syncprop(os);
1707 zilog->zl_next_batch = 1;
1709 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1711 for (int i = 0; i < TXG_SIZE; i++) {
1712 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1713 MUTEX_DEFAULT, NULL);
1716 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1717 offsetof(lwb_t, lwb_node));
1719 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1720 offsetof(itx_t, itx_node));
1722 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1724 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1725 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1727 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1728 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1729 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1730 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1736 zil_free(zilog_t *zilog)
1738 zilog->zl_stop_sync = 1;
1740 ASSERT0(zilog->zl_suspend);
1741 ASSERT0(zilog->zl_suspending);
1743 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1744 list_destroy(&zilog->zl_lwb_list);
1746 avl_destroy(&zilog->zl_vdev_tree);
1747 mutex_destroy(&zilog->zl_vdev_lock);
1749 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1750 list_destroy(&zilog->zl_itx_commit_list);
1752 for (int i = 0; i < TXG_SIZE; i++) {
1754 * It's possible for an itx to be generated that doesn't dirty
1755 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1756 * callback to remove the entry. We remove those here.
1758 * Also free up the ziltest itxs.
1760 if (zilog->zl_itxg[i].itxg_itxs)
1761 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1762 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1765 mutex_destroy(&zilog->zl_lock);
1767 cv_destroy(&zilog->zl_cv_writer);
1768 cv_destroy(&zilog->zl_cv_suspend);
1769 cv_destroy(&zilog->zl_cv_batch[0]);
1770 cv_destroy(&zilog->zl_cv_batch[1]);
1772 kmem_free(zilog, sizeof (zilog_t));
1776 * Open an intent log.
1779 zil_open(objset_t *os, zil_get_data_t *get_data)
1781 zilog_t *zilog = dmu_objset_zil(os);
1783 ASSERT(zilog->zl_clean_taskq == NULL);
1784 ASSERT(zilog->zl_get_data == NULL);
1785 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1787 zilog->zl_get_data = get_data;
1788 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1789 2, 2, TASKQ_PREPOPULATE);
1795 * Close an intent log.
1798 zil_close(zilog_t *zilog)
1803 zil_commit(zilog, 0); /* commit all itx */
1806 * The lwb_max_txg for the stubby lwb will reflect the last activity
1807 * for the zil. After a txg_wait_synced() on the txg we know all the
1808 * callbacks have occurred that may clean the zil. Only then can we
1809 * destroy the zl_clean_taskq.
1811 mutex_enter(&zilog->zl_lock);
1812 lwb = list_tail(&zilog->zl_lwb_list);
1814 txg = lwb->lwb_max_txg;
1815 mutex_exit(&zilog->zl_lock);
1817 txg_wait_synced(zilog->zl_dmu_pool, txg);
1818 ASSERT(!zilog_is_dirty(zilog));
1820 taskq_destroy(zilog->zl_clean_taskq);
1821 zilog->zl_clean_taskq = NULL;
1822 zilog->zl_get_data = NULL;
1825 * We should have only one LWB left on the list; remove it now.
1827 mutex_enter(&zilog->zl_lock);
1828 lwb = list_head(&zilog->zl_lwb_list);
1830 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1831 list_remove(&zilog->zl_lwb_list, lwb);
1832 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1833 kmem_cache_free(zil_lwb_cache, lwb);
1835 mutex_exit(&zilog->zl_lock);
1838 static char *suspend_tag = "zil suspending";
1841 * Suspend an intent log. While in suspended mode, we still honor
1842 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1843 * On old version pools, we suspend the log briefly when taking a
1844 * snapshot so that it will have an empty intent log.
1846 * Long holds are not really intended to be used the way we do here --
1847 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1848 * could fail. Therefore we take pains to only put a long hold if it is
1849 * actually necessary. Fortunately, it will only be necessary if the
1850 * objset is currently mounted (or the ZVOL equivalent). In that case it
1851 * will already have a long hold, so we are not really making things any worse.
1853 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1854 * zvol_state_t), and use their mechanism to prevent their hold from being
1855 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1858 * if cookiep == NULL, this does both the suspend & resume.
1859 * Otherwise, it returns with the dataset "long held", and the cookie
1860 * should be passed into zil_resume().
1863 zil_suspend(const char *osname, void **cookiep)
1867 const zil_header_t *zh;
1870 error = dmu_objset_hold(osname, suspend_tag, &os);
1873 zilog = dmu_objset_zil(os);
1875 mutex_enter(&zilog->zl_lock);
1876 zh = zilog->zl_header;
1878 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1879 mutex_exit(&zilog->zl_lock);
1880 dmu_objset_rele(os, suspend_tag);
1881 return (SET_ERROR(EBUSY));
1885 * Don't put a long hold in the cases where we can avoid it. This
1886 * is when there is no cookie so we are doing a suspend & resume
1887 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1888 * for the suspend because it's already suspended, or there's no ZIL.
1890 if (cookiep == NULL && !zilog->zl_suspending &&
1891 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1892 mutex_exit(&zilog->zl_lock);
1893 dmu_objset_rele(os, suspend_tag);
1897 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1898 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1900 zilog->zl_suspend++;
1902 if (zilog->zl_suspend > 1) {
1904 * Someone else is already suspending it.
1905 * Just wait for them to finish.
1908 while (zilog->zl_suspending)
1909 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1910 mutex_exit(&zilog->zl_lock);
1912 if (cookiep == NULL)
1920 * If there is no pointer to an on-disk block, this ZIL must not
1921 * be active (e.g. filesystem not mounted), so there's nothing
1924 if (BP_IS_HOLE(&zh->zh_log)) {
1925 ASSERT(cookiep != NULL); /* fast path already handled */
1928 mutex_exit(&zilog->zl_lock);
1932 zilog->zl_suspending = B_TRUE;
1933 mutex_exit(&zilog->zl_lock);
1935 zil_commit(zilog, 0);
1937 zil_destroy(zilog, B_FALSE);
1939 mutex_enter(&zilog->zl_lock);
1940 zilog->zl_suspending = B_FALSE;
1941 cv_broadcast(&zilog->zl_cv_suspend);
1942 mutex_exit(&zilog->zl_lock);
1944 if (cookiep == NULL)
1952 zil_resume(void *cookie)
1954 objset_t *os = cookie;
1955 zilog_t *zilog = dmu_objset_zil(os);
1957 mutex_enter(&zilog->zl_lock);
1958 ASSERT(zilog->zl_suspend != 0);
1959 zilog->zl_suspend--;
1960 mutex_exit(&zilog->zl_lock);
1961 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1962 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1965 typedef struct zil_replay_arg {
1966 zil_replay_func_t **zr_replay;
1968 boolean_t zr_byteswap;
1973 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1975 char name[MAXNAMELEN];
1977 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1979 dmu_objset_name(zilog->zl_os, name);
1981 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1982 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1983 (u_longlong_t)lr->lrc_seq,
1984 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1985 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1991 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1993 zil_replay_arg_t *zr = zra;
1994 const zil_header_t *zh = zilog->zl_header;
1995 uint64_t reclen = lr->lrc_reclen;
1996 uint64_t txtype = lr->lrc_txtype;
1999 zilog->zl_replaying_seq = lr->lrc_seq;
2001 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2004 if (lr->lrc_txg < claim_txg) /* already committed */
2007 /* Strip case-insensitive bit, still present in log record */
2010 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2011 return (zil_replay_error(zilog, lr, EINVAL));
2014 * If this record type can be logged out of order, the object
2015 * (lr_foid) may no longer exist. That's legitimate, not an error.
2017 if (TX_OOO(txtype)) {
2018 error = dmu_object_info(zilog->zl_os,
2019 ((lr_ooo_t *)lr)->lr_foid, NULL);
2020 if (error == ENOENT || error == EEXIST)
2025 * Make a copy of the data so we can revise and extend it.
2027 bcopy(lr, zr->zr_lr, reclen);
2030 * If this is a TX_WRITE with a blkptr, suck in the data.
2032 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2033 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2034 zr->zr_lr + reclen);
2036 return (zil_replay_error(zilog, lr, error));
2040 * The log block containing this lr may have been byteswapped
2041 * so that we can easily examine common fields like lrc_txtype.
2042 * However, the log is a mix of different record types, and only the
2043 * replay vectors know how to byteswap their records. Therefore, if
2044 * the lr was byteswapped, undo it before invoking the replay vector.
2046 if (zr->zr_byteswap)
2047 byteswap_uint64_array(zr->zr_lr, reclen);
2050 * We must now do two things atomically: replay this log record,
2051 * and update the log header sequence number to reflect the fact that
2052 * we did so. At the end of each replay function the sequence number
2053 * is updated if we are in replay mode.
2055 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2058 * The DMU's dnode layer doesn't see removes until the txg
2059 * commits, so a subsequent claim can spuriously fail with
2060 * EEXIST. So if we receive any error we try syncing out
2061 * any removes then retry the transaction. Note that we
2062 * specify B_FALSE for byteswap now, so we don't do it twice.
2064 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2065 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2067 return (zil_replay_error(zilog, lr, error));
2074 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2076 zilog->zl_replay_blks++;
2082 * If this dataset has a non-empty intent log, replay it and destroy it.
2085 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2087 zilog_t *zilog = dmu_objset_zil(os);
2088 const zil_header_t *zh = zilog->zl_header;
2089 zil_replay_arg_t zr;
2091 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2092 zil_destroy(zilog, B_TRUE);
2095 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2097 zr.zr_replay = replay_func;
2099 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2100 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2103 * Wait for in-progress removes to sync before starting replay.
2105 txg_wait_synced(zilog->zl_dmu_pool, 0);
2107 zilog->zl_replay = B_TRUE;
2108 zilog->zl_replay_time = ddi_get_lbolt();
2109 ASSERT(zilog->zl_replay_blks == 0);
2110 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2112 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2114 zil_destroy(zilog, B_FALSE);
2115 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2116 zilog->zl_replay = B_FALSE;
2117 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2121 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2123 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2126 if (zilog->zl_replay) {
2127 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2128 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2129 zilog->zl_replaying_seq;
2138 zil_vdev_offline(const char *osname, void *arg)
2142 error = zil_suspend(osname, NULL);
2144 return (SET_ERROR(EEXIST));